Agentic AI Engineer

Booz Allen Hamilton•Alexandria, VA

3d•$77,600 - $176,000

About The Position

Booz Allen is seeking an Agentic AI Engineer to build autonomous, goal-oriented AI systems and help shift its AI strategy from passive LLM chatbots to proactive, multi-agent orchestration. The role involves designing agent architectures that reason, plan, and take action; developing systems for agent communication and collaborative task solving; and building advanced retrieval-augmented generation (RAG) pipelines that combine unstructured data with Knowledge Graphs. The engineer will also fine-tune Small Language Models (SLMs) for domain-specific use and edge deployment, and develop evaluation frameworks for agent reliability, safety, and performance as systems move from prototype to production. The position calls for production-grade machine learning and software engineering experience, familiarity with agent frameworks and protocols, cloud architecture skills, and a Top Secret clearance. Booz Allen describes its benefits as including health, life, disability, financial, and retirement benefits, paid leave, professional development, tuition assistance, work-life programs, dependent care, and recognition awards. The projected annualized compensation range is $77,600 to $176,000 USD. The hiring process includes identity verification using biometrics and AI; candidates must be on camera during interviews and assessments. Use of AI or other tools to assist with interview responses is prohibited unless explicitly permitted.

Requirements

At least 2 years of software development experience.
At least 2 years of experience as an ML engineer building production-grade ML solutions using tools such as Docker or Kubernetes, including GenAI, LLMs, deep learning, reinforcement learning, AI agents, agentic workflows, or complex automation frameworks.
At least 2 years of experience with LangChain, LangGraph, AutoGen, PydanticAI, CrewAI, or LlamaIndex.
At least 2 years of experience working in cloud environments, including AWS and Azure, evaluating architectural tradeoffs, and designing robust service-based applications for scalable use.
Experience with MCP for tool integration and A2A for agent collaboration, and with RAG architecture and Knowledge Graphs, including Neo4j or NebulaGraph.
Experience fine-tuning LLMs or SLMs using Hugging Face, PEFT, or LoRA; evaluating LLM performance and behavior; and building stakeholder observation layers using tools such as Grafana, Langfuse, LangSmith, or Phoenix.
Knowledge of modern software design patterns, including microservices or edge computing.
Ability to adapt to a rapidly changing environment and navigate ambiguity.
Top Secret clearance.
Bachelor’s degree.

Nice To Haves

Experience deploying agentic systems in production.
Experience deploying agents on edge devices such as Android or using local models.
Experience integrating coding agents such as Cursor or Windsurf into a development pipeline with measured results.
Programming experience, including ML frameworks such as TensorFlow, PyTorch, llama.cpp, and vLLM.
Experience engineering AI capabilities in on-premise or multi-classification environments.
Excellent verbal and written communication skills for client engagements, client-facing project work, and business development.
Master’s degree in a computer science or AI field preferred; doctorate in computer science or statistics is a plus.

Responsibilities

Design and implement intelligent agent architectures that reason, plan, and take actions using frameworks such as LangChain, LangGraph, and AutoGen.
Develop multi-agent systems using MCP and A2A protocols to support communication, tool use, and collaborative task solving.
Build RAG pipelines that integrate unstructured data with Knowledge Graphs to improve reasoning accuracy and context retention.
Fine-tune SLMs for specific domains and optimize them for edge-device performance, including with ONNX, GGML, or Ollama.
Develop evaluation frameworks to test agent reliability, safety, and performance as systems move from prototype to production, including ReAct loops and human-in-the-loop approaches.